TPI-MINN-00/44

UMN-TH-1920/00
hep-th/0009131

Quark-Hadron Duality 1
arXiv:hep-ph/0009131v1 11 Sep 2000

M. Shifman
Theoretical Physics Institute, University of Minnesota, Minneapolis, MN 55455

Abstract

I review the notion of the quark–hadron duality from the modern perspective.
Both, the theoretical foundation and practical applications are discussed. The
proper theoretical framework in which the problem can be formulated and treated
is Wilson’s operator product expansion (OPE). Two models developed for the de-
scription of duality violations are considered in some detail: one is instanton-based,
another resonance-based. The mechanisms they represent are complementary. Al-
though both models are rather primitive (their largest virtue is their simplicity) they
hopefully capture important features of the phenomenon. Being open for improve-
ments, they can be used “as is” for orientation in the studies of duality violations
in the processes of practical interest.

1
Based on the talks delivered at the VIII-th International Symposium on Heavy Flavor Physics,
Southampton, UK, 25–29 July 1999, and the International Workshop Gribov–70, Orsay, France,
27–29 March 2000. To be published in the Boris Ioffe Festschrift At the Frontier of Particle
Physics/Handbook of QCD, Ed. M. Shifman (World Scientific, Singapore, 2001).
1 Introduction
Quantum chromodynamics (QCD) is a very strange theory. All theoretical calcula-
tions are done in terms of quarks and gluons. At the same time, quarks and gluons
are never detected experimentally. What is actually produced and detected in exper-
imental devices are hadrons: pions, kaons, protons, etc. The quark-hadron duality
allows one, under certain circumstances, to bridge the gap between the theoretical
predictions and experimentally observable quantities. The idea was first formulated
at the dawn of the QCD era by Poggio, Quinn and Weinberg [1], who suggested
that certain inclusive hadronic cross sections at high energies, being appropriately
averaged over an energy range, had to (approximately) coincide with the cross sec-
tions one could calculate in the quark-gluon perturbation theory. The uses of this
theoretical construction are countless: the e+ e− annihilation, deep inelastic scatter-
ing, hadronic τ decays, inclusive decays of heavy quarks, physics at the Z peak, to
name just a few. In spite of the dramatic developments in QCD in the subsequent
two decades the notion of the quark-hadron duality remained vague, essentially at
the 1976 level, with the very basic questions unanswered. These questions are:

• What energy is considered to be high enough for the quark-hadron duality to

set in, and what accuracy is to be expected?
• What weight function is appropriate for the averaging of the experimental cross
sections?
• If the theoretical prediction includes only perturbation theory, should one limit
oneself to some particular order in the the αs series?
• Do we have to include known nonperturbative effects (e.g. condensates) in the
theoretical prediction?
• Given a definition of the quark-hadron duality, can one estimate deviations
from duality and how?

Of course, not all of these questions are independent. For instance, answering
the last question, one will simultaneously learn the boundary energy.
Systematic explorations of these and related issues started in earnest in 1994 [2].
As in many other instances, this was dictated by practical needs. Previously, the
accuracy of the experimental data on hard inclusive processes was rather modest, so
that the Poggio-Quinn-Weinberg prescription was good enough. By 1994 the data,
mostly associated with the b quark physics, became so precise and the questions
raised so acute, that a much better theoretical understanding became imperative.
Probably, the most clear-cut example is the problem of the B semileptonic branching
ratio [3]: theoretical expectations obtained in the quark-gluon language exceed the
measured number by 10 to 20%. Possible deviation from duality is suspected to be
a major source of theoretical uncertainties. If one could reliably rule out duality

1
violations at this level, the conclusion of the interference of new physics would ensue
(provided the experimental data stay intact, of course). The stakes are quite high.
It is fair to say that (short of the full solution of QCD) understanding and
controlling the accuracy of the quark-hadron duality is one of the most important
and challenging problems for the QCD practitioners today. In this issue one cannot
expect help from lattices. The duality violation is a phenomenon inseparable from
the Minkowskian kinematics; numerical Euclidean approaches, such as lattice QCD,
have nothing to say on this issue. Analytic methods are needed.
In this review I will summarize the results of an investigation [4]–[10], which spans
over six years and is not yet complete. I will discuss the modern formulation of the
problem based on Wilson’s operator product expansion (OPE). This formulation is
solid and unambiguous. Then I will review models which were designed to give us
a certain idea of (and a degree of control over) the deviations from duality. There
are two classes of such models: instanton-based and resonance-based. They are
complementary to each other; I will discuss both classes. Finally, I will present
sample applications in the processes of the current interest, such as the hadronic τ
decays.

2 The Quark-Hadron Duality: What Does That

Mean?
Let us consider an idealized theory – QCD, with two massless quarks, u and d.
We are interested in the total hadronic cross section of the e+ e− annihilation. The
“photon” in our theory is idealized too, its coupling to the quark current has the
form
¯ µd
ūγµ u − dγ
Jµ = √ . (1)
2
(In fact, this is the isovector part of the actual electromagnetic current). We define
the two-point function Πµν
Z
Πµν = i eiqx d4 xh0|T {Jµ(x)Jν† (0)}|0i . (2)

Here q is the total momentum of the quark-antiquark pair. Due to the current
conservation Πµν is transversal,

Πµν = (qµ qν − q 2 gµν )Π(q 2 ) . (3)

The experimentally observable quantity is the imaginary part of Π(q 2 ) at positive

values of q 2 (i.e. above the physical threshold of the hadron production), the spectral
density,
12π
ρ(s) = Im Π(s) , s ≡ q 2 . (4)
Nc

2
 q q

Figure 1: The one-loop graph determining the polarization operator and the spectral
density in the leading (parton) approximation. The “photon” momentum is denoted
by q.

Up to a normalization, the expression above coincides with the cross section of e+ e−

annihilation into hadrons measured in the units σ(e+ e− → µ+ µ− ), the famous ratio
R.
Theoretically one can calculate Π(q 2 ) in the deep Euclidean domain, at negative
2
q . For instance, from the free quark loop of Fig. 1 one gets
Nc
Π(Q2 ) → − ln Q2 Q2 ≡ −q 2 . (5)
12π 2
Performing the analytic continuation to the Minkowski domain and taking the imag-
inary part one arrives at
ρ(s)theor → 1 , s → ∞ . (6)
This is the spectral density in the theory with free quarks, i.e. αs = 0. There
are various corrections to the free quark result (5). The perturbative gluon ex-
changes give rise to the αs (Q2 ) series. Nonperturbative (power) corrections come
from the quark and gluon condensates and from other sources, e.g. the small-size
instantons. A systematic method of handling the theoretical calculations of Π(Q2 )
in the deep Euclidean domain is provided by Wilson’s operator product expansion
(OPE) [11]. Essentially, this is a bookkeeping procedure: one consistently separates
the short-distance contributions (i.e. those coming from distances ≤ µ−1 ) from the
large distance contributions (i.e. those coming from distances ≥ µ−1 ). Here µ is
a theoretical parameter (usually referred to as the normalization point) separating
the two domains. The choice of µ is a matter of convenience – observable quantities
do not depend on it.
The short-distance contributions determine the coefficients Cn (q) of OPE ,

D(q 2 ) = D(Q2 ) ≡ −(4π 2 )Q2 (dΠ/dQ2 ) .

X
Cn (q; µ)hOn (µ)i , (7)
n

The normalization point µ is indicated explicitly. The sum in Eq. (7) runs over all
possible Lorentz and gauge invariant local operators built from the gluon and quark
fields. The operator of the lowest (zero) dimension is the unit operator I, followed
by the gluon condensate G2µν , of dimension four. The four-quark condensate gives
an example of dimension-six operators.
At short distances QCD is well described by the quark-gluon perturbation theory.
Therefore, as a first approximation, it is reasonable to calculate Cn perturbatively,

3
in the form of expansion in αs (Q) ∼ (ln Q)−1 . This certainly does not mean that
the coefficients Cn are free from nonperturbative (nonlogarithmic) terms. The latter
may and do appear in Cn ’s; they are of the type ∼ Q−γ where γ is a positive number,
not necessarily an integer. Such terms in Cn ’s are generated, for instance, by the
small-size instantons. Another source of the power terms in Cn , of a technical rather
than dynamical nature, is the normalization point µ: in calculating Cn (µ) one must
remove all soft contributions with off-shellness less than µ.
The condensate terms in Eq. (7) give rise to corrections of the type (ΛQCD /Q)n
where n is an integer ≥ 4, the (normal) dimension of the operator On , modulo
logarithms associated with the anomalous dimensions. ΛQCD is the scale parameter
of QCD (sometimes, I will drop the subscript “QCD” for brevity) entering through
the matrix elements of On ’s.
If one could calculate Π(Q2 ) in the Euclidean domain exactly, one could analyt-
ically continue the result to the Minkowski domain, and then take the imaginary
part. The spectral density ρ(s)theor obtained in this way would present the exact
theoretical prediction for the measurable hadronic cross section. There would be no
need for duality.
In practice, our calculation of Π(Q2 ) is approximate, for many reasons. First,
nobody is able to calculate the infinite αs (Q2 ) series for the coefficient functions, let
alone the infinite condensate series. Both have to be truncated at some finite order.
A few lowest-dimension condensates that can be captured, are known approximately.
The best we can do is analytically continue the truncated theoretical expression, term
by term, from positive to negative Q2 . For each term in the expansion the imaginary
part at positive q 2 (negative Q2 ) is well-defined. We assemble them together and
declare the corresponding ρ(s)theor to be dual to the hadronic cross section ρ(s)exp .
In the given context “dual” means equal.
Let me elucidate this point in more detail. Assume that Π(Q2 ) is calculated
through αs2 and 1/Q4 , while the terms αs3 and 1/Q6 (with possible logarithms) are
dropped. Then the theoretical quark-gluon spectral density, obtained as described
above, is expected to coincide with ρ(s)exp , with the uncertainty of order O[(αs (s))3 ]
and O(1/s3). The uncertainty in the theoretical prediction of this order of magnitude
is natural since terms of this order are neglected in the theoretical calculation of
Π(Q2 ). If the coincidence in this corridor does take place, we say that the quark-
gluon prediction is dual to the hadronic spectral density. If there are deviations
going beyond the natural uncertainty, we call them violations of duality. Needless to
say that, once our calculation of Π(Q2 ) becomes more precise, the definition of the
“natural uncertainty” in ρ(s)theor changes accordingly.
This is the most clear-cut definition I can suggest. From the formal standpoint,
it connects the duality violation issue with that of analytic continuation from the
Euclidean to Minkowski domain. Negligibly small corrections (legitimately) omitted
in the Euclidean calculations may and do get enhanced in Minkowski.
Before I proceed to explain, from the physical standpoint, where the violations
of duality come from, and their pattern, I would like to make a remark important in

4
the conceptual aspect. The necessity of truncation of the αs and condensate series
is not just due to our practical inability to calculate high-order terms. (For the
vast majority of the theorists, myself including, “high-order” begins at the next-
to-next-to-leading level.) Assume we have a “sorcerer’s stone” which would allow
us to exactly calculate any term in the expansion we want. Still, we would not be
able to find Π(Q2 )exact because the both series are factorially divergent. The αs
series has at least two known sources of the factorial behavior: first, the number of
the Feynman graphs grows factorially at high orders [12]; second, there are graphs
with renormalons, (of which we will not bother in what follows since the infrared
renormalons – the only ones which are potentially dangerous – are totally eliminated
by the introduction of the normalization point µ [13]). The condensate series is
divergent too. The factorial divergence of the condensate series is studied even to
a lesser extent than that of the αs series. The only fact considered to be firmly
established [2, 4] is the divergence per se. It is not Borel-summable, which sets a
limit of the theoretical accuracy. Including more and more terms in the series would
not help, and even an optimal truncation would leave a gap between Π(Q2 )theor and
Π(Q2 )exact .

3 Where the Duality Violations Come From?

Theoretical calculations of inclusive processes in QCD are performed through Wil-
son’s OPE in the Euclidean domain. Therefore, to understand what is included in
such calculations and what is left out, one should have a clear picture of limitations
of OPE.
Let us return to the consideration of the correlation function of two currents,

h0|T {Jµ(x)Jν† (0)}|0i ,

which determines the polarization operator in Eq. (2), at negative (Euclidean) x2 .

Wilson’s OPE is nothing but an expansion of h0|T {Jµ (x)Jν† (0)}|0i in singularities at
x2 = 0. It properly captures all terms of the type 1/x6 times logarithms, or ln x2 , or
x2 ln x2 , and so on. Every term of this type is represented in the condensate series
provided that the calculation is carried out to a sufficient order.
Let me note in passing that limiting oneself to singular terms (and, in particular,
to the leading singular term) is a reasonable approximation in the theories in which
the asymptotic conformal regime sets in at short distances in the power-like manner.
In fact, OPE in its original formulation was designed by Wilson for applications
in such theories (see the first work in Ref. [11]). In QCD the approach to the
asymptotically free regime is very slow – logarithmic. That’s why the precision of
the leading (parton) approximation is sufficient only for very rough estimates, and
that’s why the high-order terms, both logarithmic and power, must be kept. The
time when the QCD practitioners would be satisfied by the leading, or even the first
subleading approximation, is long gone.

5
 It is clear that the function h0|T {Jµ(x)Jν† (0)}|0i is not fully determined by its
singularities at x2 = 0. Generally speaking, one may have, additionally, isolated
singularities at finite x2 , or a singularity at infinity, which are not reflected in the
truncated Wilson’s OPE. Consider, say, a singularity at finite x2 of the form
1
. (8)
x2 + ρ2
The expansion of this function (truncated at any order) generates derivatives of
δ(q 2 ) in the momentum space; therefore, it has no impact on the OPE expansion
at large Q2 . This contribution is clearly missing. The Fourier transform of Eq. (8)
at large Euclidean Q2 falls off as exp(−Qρ). Thus, this term is smaller than any
of the terms in the condensate expansion. We should not forget, however, that our
final goal is a prediction in the Minkowski domain. Upon analytic continuation, the
exponentially small term exp(−Qρ) looses its suppression √ and becomes oscillating,
sin(−Eρ), where E stands for the total energy, E = q 2 . If it were not for the
power suppression in the pre-exponent, such exponential/oscillating terms would be
a total disaster. Since they are not seen in OPE, any prediction for the inclusive cross
sections made through OPE (which is equivalent, in practice, to perturbation theory
plus a few condensates) would be grossly wrong. There would be no quark-hadron
duality at all. All calculations of the hard processes – from the total hadronic cross
sections in e+ e− annihilation, to jet physics, to the heavy quark decays – would be
valid roughly up to a factor of two, no matter how large the energy release is.
Fortunately, one avoids the disaster due to the fact that the duality-violating
exponential/oscillating terms are, in fact, suppressed as E −κ sin(−Eρ) where κ is
a positive index, which depends on the process under consideration and is typically
rather large. This justifies theoretical predictions based on a few first terms in
the condensate expansion. Needless to say that determining κ is one of the most
important tasks in the issue of the quark-hadron duality.
Singularities at x2 → ∞ also lead to the exponential/oscillating terms of a
somewhat different form, of the type exp(−Q2 ρ2 ) in Euclidean. In the Minkowski
domain one gets an oscillating function of the type E −η sin(−E 2 ρ2 ), where η is
another positive index. In one and the same process, one can expect both duality-
violating components to show up. Generally speaking, the indices κ and η are
unrelated; at least, at the moment we do not see any obvious relation between
them. (The coincidence of κ and η in the practically important problem of the
inclusive hadronic τ decays, κ = η = 6, seems to be accidental, see Sec. 8).
How can one isolate the singularities at x2 = 0 from others diagrammatically?
To answer this question it is convenient to pass to the momentum space. In the
leading approximation the polarization operator Π(Q2 ) is presented by the graph of
Fig. 1. The large Euclidean momentum q (remember, Q2 → ∞) flows in the photon
line on the left, propagates through the fermion lines, and leaves the graph through
the photon line on the right. The virtual momenta of the fermion lines typically
scale as q. The first αs correction is presented in Fig. 2. In this graph the virtual

6
 q q
k

Figure 2: The one-gluon correction in the the polarization operator. The gluon
momentum is denoted by k.

momenta of all lines in the loops, the fermion and the gluon, scale as q. Thus, the
diagram of Fig. 2 determines the leading logarithmic correction to the parton result
(5) or (6).
Although the leading contribution to the integral comes from the domain where
all virtual momenta are proportional to q, there are nonvanishing contributions
from other kinematical domains. For instance, one can consider a “corner” where
the gluon virtual momenta k in Fig. 2 is small, |k| < µ, and does not scale with q.
Certainly, at small k the gluon propagator is not given by perturbation theory. The
gluon line must be cut. This corresponds to the gluon condensate term [14] in the
condensate expansion.
In general, any term in the condensate expansion can be interpreted in this way,
in terms of factorization in the momentum space. The large external momentum q is
transmitted through one or several hard lines – their momenta scale with q. This is
the definition of “hardness.” The remainder of the graph is a soft part, which factors
out and gives rise to gluon, quark and mixed condensates. The virtual momenta in
this part of the graph do not scale with q, they are limited by a fixed parameter µ
(this is the definition of “softness”).
The hard part of the graph is responsible for the coefficient functions. The fact
that not all lines in the given graph are hard results in the power suppression of
the corresponding coefficient function. Letting more and more lines to be soft, one
obtains consecutive terms in the condensate expansion. What is missing?
It is conceivable that the number of lines through which q is transmitted becomes
so large, that though Q → ∞, neither of the lines is hard. Of course, in this case
one cannot speak of the separate lines in the Feynman graphs. It would be more
relevant to say that the external momentum q is transmitted from the incoming to
the outgoing photon through a coherent soft field fluctuation, see Fig. 3. An example
is provided, for instance, by a fixed-size instanton. It is clear that this mechanism
is conceptually related to the truncated tail of the condensate series. Indeed, as
one proceeds to higher condensates, more lines become soft. Eventually we arrive
at the situation when all lines are soft. Mathematically, the exponential/oscillating
contribution is related to the factorial divergence of the condensate series [4]. This
is the usual story. The exponentially small terms in Euclidean convert into an
oscillating function in Minkowski.
I will use the fixed-size instantons for the purpose of modeling this mechanism
of duality violations. The observation that “soft” instantons generate an oscillating

7
 q q

soft
field

Figure 3: Transmitting a large external momentum through a soft field.

component ascends to Refs. [15, 16]. By no means I imply that the instantons
are the dominant soft fields in the QCD vacuum. True, there are models in which
they are assumed to be dominant, the so-called instanton liquid models [17]. Their
status in the range of questions I am interested in (the duality violations) has yet
to be clarified. I will use instantons for the purpose of orientation, in hope that at
least some features of the results obtained in this way will be more general than the
model itself.

4 Model or Theory?
The topic I address – the quark-hadron duality violations – has a unique status.
By definition, one cannot build an exhaustive theory of the duality violations based
on the first principles. Indeed, assuming there is a certain dynamical mechanism
(which goes beyond perturbation theory and condensates) for which such a theory
exists, one will immediately include the corresponding component in the theoretical
calculation. The reference quantity, Π(Q2 )theor , will be redefined accordingly. After
the analytic continuation to Minkowski, this will lead, in turn, to a new theoretical
spectral density to be used as a reference ρ(s)theor in the duality relation.
Thus, by the very nature of the problem, it is bound to be treated in models
of various degrees of fundamentality and reliability. This is because the duality
violation parametrizes our ignorance. Ideally, the models one should aim at must
have a clear physical interpretation, and must be tested, in their key features, against
experimental data. This will guarantee a certain degree of confidence when these
models are applied to the estimates of the duality violations in the processes and
kinematical conditions where they had not been tested.

8
5 The Physical Picture Behind the Duality
Before delving into technical details I will describe the phenomenon from a slightly
different perspective. The quark-hadron duality takes place in those processes where
one can isolate two stages in the process under consideration, occurring at two
distinct scales. A basic transition involving quarks (gluons) must typically occur at
a short scale regulated by external parameters such as Q, mQ , etc. For instance,
in the e+ e− annihilation the basic transition is the conversion of the virtual γ into
¯ Then, at the second stage, the quarks (gluons) materialize in the form
ūu or dd.
of hadrons, at a much larger scale. In the appropriate frame, the first time scale
is of order 1/Q while the second of order Q/Λ2 . By that time, the original quarks
are far away from each other – a residual interaction cannot significantly alter the
transition cross section which was “decided” at the first (quark-gluon) stage.
The duality violations are due to (i) rare atypical events, when the basic quark
transition occurs at large rather than short distances; (ii) residual interactions oc-
curring at large distances between the quarks produced at short distances. In the
first case appropriate (Euclidean) correlation functions develop singularities at finite
x2 , while the second mechanism is correlated with the x2 → ∞ behavior.
In both cases the duality violating component follows the pattern I have dis-
cussed above – exponential in Euclidean and oscillating in Minkowski. Three distinct
regimes were identified and considered in the literature so far:
• (i) Finite-distance singularities
√
s−κ/2 sin( s) ; (9)

• (ii) Infinite-distance singularities (Nc = ∞)

s−η/2 sin(s) ; (10)

• (iii) Infinite-distance singularities (Nc large but finite, s → ∞)

1
 
exp (−αs) sin(s), α=O ≪ 1. (11)
Nc
These regimes are not mutually exclusive – in concrete processes one may expect
the duality violating component to be a combination of (i) and (ii), or (i) and (iii).
From the theoretical standpoint it is quite difficult to consistently define the duality
violating component of the type (3). An operational definition I might suggest is as
follows: Start from the limit Nc = ∞ and identify the component of the type (2).
Follow its evolution as Nc becomes large but finite.

6 An Instanton-Based Model
The basic features of the formalism [5] to be used below to model exponential/oscil-
lating terms (which present deviations from duality) are as follows:

9
 (i) One considers quarks propagating in the instanton background field. The
instanton size ρ is assumed to be fixed. Alternatively, one can say that an effective
instanton measure has a δ-function peak in ρ.
(ii) For the light quarks, most instanton amplitudes are suppressed by powers
of the light quark masses. This suppression is due to the fermion zero modes of
the Dirac operator, occurring in the instanton-like backgrounds. We ignore these
factors, as well as all other pre-exponential overall factors coming from the instanton
measure. We will only trace dependences on large momenta relevant to the problems
under consideration (Q in e+ e− annihilation, the heavy quark mass mQ in the case
of the heavy quark decays, and so on).
(iii) We ignore all singularities of the correlation functions at x = 0. The cor-
responding contributions are associated with the power terms in the condensate
expansions. The instantons-based models are presumably not precise enough to
properly capture the condensates. We will isolate and calculate only those contribu-
tions that come from the finite distance singularities at x2 = −ρ2 . In the momentum
space they produce the exponential/oscillating terms sought for.
It is seen that our evaluation of deviations from duality is based on the most
general aspects of the instanton formalism, and, in essence, does not depend on
details.
The advantage of the model is its simplicity. The only formula we will need is
ν−2
1 2π 2 Qρ K2−ν (Qρ)
Z 
d4 x eiqx
= (12)
(x2 + ρ2 )ν Γ(ν) 2 ρ2ν−4
in the Euclidean domain, which implies that in the Minkowski domain, at large q 2 ,
Z
1 ν 5 √
Im d4 x eiqx ∝ s 2 − 4 sin( sρ − δ) , s = q2 . (13)
(x2 2
+ρ ) ν

Here K is the McDonald function, and δ is a constant phase which is of no concern

to us here.
To explain how it works it seems best to consider concrete examples. Let us
start from e+ e− annihilation. The polarization operator defined in Eq. (2) is given
by the graph of Fig. 1. In Sec. 2 we evaluated this graph for free quarks and found
that ρ(s) tends to a constant at asymptotically large energies. Now, we evaluate
the very same graph using the quark Green functions in the background instanton
field, rather than the free quark Green functions. For massless quarks the Green
functions Ginst are known exactly [18], but we will not need the full expression. We
will only need to know that Ginst is a sum of terms of the following structure
1 1
Ginst (x, y) = G̃ , (14)
[(x − y)2 ]2 [(x − z)2 + ρ2 ]ℓ1 [(y − z)2 + ρ2 ]ℓ2
where z is the instanton center, and
1 3
ℓ1,2 = or , ℓ1 + ℓ2 = 2 , (15)
2 2

10
and the numerator G̃ is a polynomial of x, y. As was explained above, the singularity
of Ginst (x, y) at x − y = 0 is irrelevant for the exponential/oscillating terms. Of
importance are the singularities in the complex plane coming from the second factor
in Eq. (14). The integrals which one has to take can (and must) be evaluated at
a saddle point; a simple analysis [5] shows that at the saddle point the instanton
center z is exactly in the middle between x and y. Then the relevant singularities
of the quark correlation functions are at (x − y)2 = −4ρ2 (see Eq. (14)). At the
singularity the first factor 1/(x−y)4 (which is singular at the origin) can be replaced
by 1/(16ρ4 ) and then safely omitted together with all other prefactors.
We can proceed now to the calculation of the exponential/oscillating component
of the polarization operator Πµν . The polarization operator is the product of two
Green functions (14); therefore, at large Euclidean momenta
1
Z
Πµν ∝ d4 x eiq(x−y) d4 z
[(x − z)2 + ρ2 ] [(y − z)2 + ρ2 ]3
1 1
Z Z
4 iqx
∝ d xe 2 2
× d4 zeiqz 2
x +ρ (z + ρ2 )3
1
∝ K1 (Qρ)K−1 (Qρ) ∝ exp(−2Qρ) . (16)
Q
Note that once the integration over the instanton center is carried out, the integral
factorizes.
Equation (16) implies that the exponential component of Π(q 2 ), defined in Eq.
(3), is
1
∆Π ∝ 3 exp(−2Qρ) , (17)
Q
√
which implies, in turn, that at large E = s the oscillating component of the
spectral density is
1
∆ρ(s) ∝ 3 sin(2Eρ) . (18)
E
Equation (18) reproduces the high-energy asymptotics of the exact result [19] for
the polarization operator in the one-instanton approximation.
Does the s−3/2 fall off of the oscillating (duality violating) component make
sense? I will confront theoretical expectations with experiment in Sec. 11. Here I
just refer the reader to Fig. 11 (see the dashed curve), postponing a more detailed
discussion till after both, the instanton-based and the resonance-based models, are
considered.
The next example to be analyzed is the total hadronic τ width. This exercise is
quite similar to previous exercises. The corresponding transition operator is depicted
in Fig. 4 (its imaginary part at p2 = Mτ2 is proportional to the width of the hadronic
τ decay). The neutrino Green function clearly does not “feel” the background gluon
field, so that the neutrino propagator is that of a free fermion. The same is valid

11
 q

τ ν τ

Figure 4: The transition operator T̂ determining the hadronic τ width, Γhadr (τ ) =

Mτ−1 Im hτ |T̂ |τ i.

for the τ lines. Therefore, we immediately conclude that

1
∆Γ(τ → ν + hadr.) ∝ sin(2Mτ ρ) , (19)
Mτ
cf. Eq. (16). The asymptotic (parton-model) prediction in this case is Γ(τ →
ν + hadr.) ∝ Mτ5 , so that the oscillating component in Rτ scales as
1
∆Rτ ∝ sin(2Mτ ρ) , (20)
Mτ6
where
Γ(τ − → ντ + hadrons)
Rτ ≡ . (21)
Γ(τ − → ντ e− ν̄e )

Note that the pre-exponential suppression factor in ∆Rτ is significantly stronger

than in ∆R(e+ e− ), namely, Mτ−6 vs. E −3 . Of course, to make quantitative state-
ments it is not sufficient to establish the scaling laws; one needs absolute normal-
izations. Here, we are basically in uncharted waters. At best, we have [5] some
educated guesses, which, if true, imply that ∆Rτ /Rτ < ∼ 5%. This estimate is rather
close to what one obtains for duality violations in τ in the resonance-based model,
see Sec. 8.
In Sec. 1 I mentioned that the current interest to the problem of the quark-
hadron duality was driven, to a large extent, by a significant progress in the experi-
ments on the inclusive heavy flavor decays. Theoretically, they can be treated along
the same lines as e+ e− annihilation or τ decays. The distinctions are technical. Let
us start from the total semi-leptonic width of the b flavored hadrons. At the quark
level the process is described by the transition

b → qℓ− ν ,

where q is either u or c quark, and ℓ stands for the electron, muon or τ lepton. We
will first neglect the masses of the final fermions; this is an excellent approximation
for q = u and ℓ = e, µ. The impact of the final quark (lepton) masses will be
considered later.
The relevant transition operator is depicted in Fig. 5. The differences compared

12
 q

b ν b

Figure 5: The transition operator T̂ determining the total semileptonic width of B

mesons.

to the case of the τ decay are as follows:

(i) There is one (rather than two) light-quark line carrying a large external
momentum;
(ii) Unlike the τ lines, the b quark lines do experience interactions with the
background gluon field.
The analysis of the exponential/oscillating component in the heavy flavor decays
combines standard elements of the heavy quark expansion (for a review see e.g.
[20]), and the instanton calculus (a pedagogical review can be found in [21]). It is
convenient to choose the rest frame of the heavy hadron at hand, and single out the
large “mechanical” part in the x dependence of the heavy quark field,

Q(x) = e−imQ t Q̃(x) .

Then the total width is proportional to the imaginary part of the transition operator,
1
Γ= hHQ |T̂ |HQ i , (22)
MHQ

where
¯
Z
T̂ = i Q̃(x)G(x, y)Q̃(y)D(x − y)eimQ (x0 −y0 ) d4 (x − y)d4 z , (23)

G(x, y) is the light quark Green function, D(x − y) describes the propagation of
colorless objects (the lepton pair in the case at hand), while z is the instanton
center. The subscript 0 marks the time component.
Both, the light quark Green function G(x, y) and the heavy quark fields Q̃ are to
be considered in the background gluon field. Taken separately, they are not gauge
invariant; the product in Eq. (23) is. In the leading order in the heavy quark
expansion, the heavy quarks propagate only in time; therefore,
R x0
Q̃(x0 , ~x) = T ei 0
A0 (τ,~
x)dτ
Q̃(0, ~x) ≡ U(x)Q̃(0, ~x) . (24)

It is convenient (although not necessary) to impose the condition that at large dis-
tances from the instanton center the quark propagation becomes free. This condition
implies that the singular gauge is used. An explicit expression for U(x) can be found
in this gauge; it is rather cumbersome, and I will not quote it here, referring the
reader to the original publication [5]. At the saddle point (which again corresponds

13
to the instanton situated exactly in the middle between the points, x and y, i.e.
z = (1/2)(x + y)) the product U −1 (x)...U(y) in Eq. (23) reduces to unity, at least
in the part which is singular at (x − y)2 = −4ρ2 . Note, that the matrices converting
the nonsingular-gauge Green function G(x, y) to the singular gauge can be ignored
too. This implies that one may continue to use Eq. (14).
Thus, the heavy quarks decouple from the instanton background in the calcula-
tion of the exponential/oscillating terms. This is the consequence of the fact that we
exploit the heavy quark expansion and limit ourselves to the leading in 1/mQ terms.
In the subleading terms this decoupling does not necessarily take place. Generally
speaking, the replacement of the z integral by the value of the integrand at the
saddle point is not warranted in the next-to-leading orders. This effect will not be
further pursued, however.
Summarizing, the heavy quark expansion makes the heavy quarks sterile with
respect to the duality violating component. This is certainly not counterintuitive.
The duality violating component in the heavy quark inclusive decays originates from
the light-quark propagators. Its calculation reduces to Eq. (14) and the routine
outlined after this equation (plus the basic formula (12)). Starting from Eq. (23)
we arrive at
1
T̂ ∝ 3 e−2mQ ρ (25)
mQ
in the Euclidean domain, which results in
1
∆Γsl ∝ sin(2mQ ρ) . (26)
m3Q

Since the total width scales as Γ ∝ m5Q , the oscillating component of the semileptonic
branching ratio is obviously suppressed by m8Q ,

1
∆Br(HQ → ℓν + hadrons) ∝ sin(2mQ ρ) . (27)
m8Q

It is easy to generalize this formula to include an arbitrary number of the light

quarks in the final state. Each extra light quark adds two powers of mQ in the
numerator. Thus, in the total nonleptonic branching ratio, with three light quarks
in the final states,
1
∆Br(HQ → light hadrons) ∝ sin(2mQ ρ) . (28)
m4Q

For the sake of completeness I will also mention the radiative decays of the type
b → s + γ. Assuming that these decays are induced by local operators2

b̄σµν (1 + γ5 )sFµν ,
2
In actuality this is true only for a part of the amplitude.

14
where Fµν is the photon field strength tensor, we immediately conclude that

1
∆Γ(b → s + γ) ∝ sin(2mQ ρ) , (29)
m3Q

precisely in the same way as in Eq. (26). The parton expression for Γ(b → s + γ)
scales as
Γ0 (b → s + γ) ∝ m3Q , mQ → ∞ . (30)
As a result, our instanton-based model predicts that the duality violating component
in b → s + γ is suppressed as
1
∆Γ(b → s + γ)/Γ0 (b → s + γ) ∝ sin(2mQ ρ) . (31)
m6Q

So far it was assumed that the quarks and leptons produced were massless. What
changes would one decide to take into account finite (nonvanishing) masses of the
quarks and leptons?
The exact dependence on the masses of the final quarks or leptons is rather
sophisticated. Although it is calculable in principle, the corresponding calculations
are much harder to perform than the simple estimates presented above. Moreover,
this is hardly necessary. Given a crude nature of the model, which is intended for
orientation rather than for precision estimates, it seems reasonable to treat u, d and
s quarks as massless while c as heavy, and the c lines as free propagation. Then the
only impact of the c quark mass is kinematical, it results in the replacement of the
total energy mQ by a relevant energy release in the light quarks in the process at
hand.

⋆⋆⋆

These are just a few of important applications which are under discussion in
the current literature. The results are collected in Table 1. Our instanton-based
model of the duality violation is user-friendly – it is very easy to evaluate the expo-
nential/oscillating component in other inclusive processes not included in Table 1,
would such a necessity arise.

7 Global vs. Local Duality

Usually by local duality people mean point-by-point comparison of ρ(s)theor and
ρ(s)exp , while global duality compares the spectral densities ρ(s) averaged over some
ad hoc interval of s, with an ad hoc weight function w(s),
Z s2 Z s2
ds w(s) ρ(s)theor ≈ ds w(s) ρ(s)exp .
s1 s1

15
 e+ e− → X τ → νX HQ → ℓνX HQ → X HQ → Xγ HQ → HQ ′ X

1 1 1 1 1 1
E3 Mτ6 8
MH 4
MH 6
MH 6
Q Q Q (∆MHQ )

Table 1: The index of the power fall-off of the oscillating (duality violating) com-
ponent in various inclusive processes, normalized to the corresponding asymptotic
(parton model) formulae, in the instanton-based model. The capital X denotes the
light-quark hadronic states.

Here I would like to comment on a very common misconception which travels

from one paper to another. Many authors believe that global duality defined in this
way has a more solid status than local duality. Some authors go so far as to say
that while global duality is certainly valid at high energies, this is not necessarily
the case for local duality. This became a routine statement in the literature. Well,
routine does not mean correct.
In fact, both procedures have exactly the same theoretical status. The point-
by-point comparison, as well as the comparison of ρ(s)’s (with an ad hoc weight
function), must be considered as distinct versions of local duality. The distinction
between the “local” quantities, such as R(e+ e− ) at a certain value of s and the
integrals of the type involved, say, in Rτ is quantitative rather than qualitative.
Comparison of Eqs. (18) and (20) makes this assertion absolutely transparent: in
both quantities there is a duality violating component, the only distinction is a
concrete index of the power fall-off (3 vs. 6).
The genuine global duality applies only to special integrals which can be directly
expressed through the Euclidean quantities. For instance, if the integration interval
extends from zero to infinity, and the weight function is exponential, the integral
Z ∞
ds exp{−s/M 2 }ρ(s) ,
0

reduces [14] to the Borel transform of the polarization operator Π(Q2 ) in the Eu-
clidean domain (i.e. at positive Q2 ). For such quantities, duality cannot be violated,
by definition.
There is one more aspect of averaging (smearing) which is not fully understood in
the literature and requires comment. Let us consider again Rτ . It can be expressed
in terms of spectral densities ρV and ρA in the vector and axial-vector channels,
respectively,
!2 !
Z Mτ2 ds s s
Rτ ≡= 2
1− 2 1 + 2 2 [ρV (s) + ρA (s)] . (32)
0 Mτ Mτ Mτ
(More exactly, the integration over s runs from Mπ2 , but in the chiral limit we stick
to, the pion mass vanishes.) The spectral densities ρV and ρA are normalized in

16
such a way that their asymptotic (free-quark) values are

ρV (s) , ρA (s) → Nc at s → ∞ .

Correspondingly, the asymptotic value of Rτ is Rτ0 = Nc .

In the instanton-based model, the duality violating component in ρV,A scales as
−3
E . It is tempting to say then, that the duality violating component in Rτ can
be obtained by integrating the duality violating components of ρV,A with the weight
function specified in Eq. (32).
This would lead us nowhere, however. First of all, the integral in Eq. (32)
runs all the way down to zero, while Eq. (18) is valid at asymptotically large E.
Even if the lower limit of integration were chosen to be high enough, this would
not help to find ∆Rτ from Eq. (32). Indeed, ∆ρV,A is an oscillating function of s,
any smearing inevitably entails cancelations, and the result would depend on subtle
details of ρ(s), which are certainly beyond theoretical control. The cancelations are
seen, in particular, from the fact that asymptotically ∆Rτ (Mτ ) falls off faster than
∆ρV,A (s), namely, Mτ−6 versus E −3 . There is no way one could predict the change
of the index from 6 to 3 based solely on the integral (32).
At the same time, our instanton-based model does allow one to predict the index
in Rτ . To this end one must analyze the appropriate transition amplitude in the
Euclidean domain performing the analytic continuation to the Minkowski domain
at the very end. The model captures enough intricate features of QCD to “know”
the result of the smearing as a whole. Note that ∆Rτ (Mτ ) depends on the highest
scale in the problem at hand, Mτ , rather than on any intermediate or low scales one
encounters in process of integration in Eq. (32). This is a general feature which will
be valid in any process and will persist in any model based on evaluating singularities
off the origin.
The case of Rτ is quite typical. Similar questions arise in other problems, where a
similar strategy should be applied. A problem of this type which deserves a special
mention in view of its practical significance, is that of the inclusive semileptonic
decays of HQ . The decay rate can be obtained as an integral over appropriate
kinematic variables over the hadronic structure functions,
G2F
Z
2
Γ(HQ → ℓν + X) = |VqQ | dEℓ dq 2 dq0
64π 2
n h i o
2q 2 w1 + 4Eℓ (q0 − Eℓ ) − q 2 w2 + 2q 2 (2Eℓ − q0 )w3 (33)

where q is the lepton pair momentum, w1,2,3 are the hadronic structure functions
which depend on q0 and q 2 (for their definition see e.g. Ref. [22]). Equation
(33) refers to the rest frame of HQ . Other notations are self-explanatory. The
exponential/oscillating contribution to the decay rate is presented in Eqs. (25) and
(26), where it is assumed that mq = 0. If one decided to get an estimate from Eq.
(33), by integrating the duality violating contributions in the structure functions,
one would observe parametrically larger violations, exploding at the boundaries of

17
the kinematically allowed domain. These large violations cancel in the total rate
∆Γ(HQ → ℓν + X), because of oscillations.

8 A Resonance-Based Model
Now we will acquaint ourselves with another approach based on the resonance sat-
uration of the colorless n-point functions. That this is a distinct dynamical source
of duality violations follows from the discussion in Secs. 3 and 5. Shortly we will
confirm this by observing that the functional form of oscillations comes out differ-
ent compared to that in the instanton-based model. Theoretical analysis is most
transparent in the limit Nc = ∞. Later I will allow Nc to become finite albeit
large (Sec. 10). We will first try to abstract general features, and then illustrate
them in the simplest possible dynamical setting which still exhibits confinement –
two-dimensional ’t Hooft model.[23]
Let us first summarize what we expect to get for the polarization operator defined
in Eq. (2) in the limit of large Nc . In multicolor QCD, Nc → ∞, the resonance
widths vanish. The spectrum of excitations in the given channel is expected be
(asymptotically) equidistant. A string-like picture of the color confinement naturally
leads to (approximately) linear Regge trajectories. For each primary trajectory
there are infinitely many daughter ones. The daughter trajectories are parallel to
the primary trajectory and are shifted by integers (for a review see e.g. [24]). As a
result, the excitation spectrum in the given channel takes the form

Mn2 = M02 + σ 2 n , σ 2 ≡ 2/α′ , (34)

where α′ is the slope of the trajectories. Note that the neighboring resonance states
in the given channel are separated by the interval 2/α′ rather than 1/α′. This is due
to the alternating signatures of the daughter trajectories.
All these properties can be extracted from the Veneziano amplitude found in
1968 (see Sec. 7.4 in Collins, [24]) which gave rise to the modern string theory.
Equation (34) implies, in turn, that Π(q 2 ) can be presented as an infinite sum
∞ ∞
Nc σ 2 X 1 Nc σ 2 X 1
Π(q 2 ) = − = − . (35)
2 2 2
12π n=0 q − Mn 12π n=0 q − σ n − M02
2 2 2

In the problem at hand there are two large quantities: the number of colors and
the energy. If one fixes s and lets Nc grow, one eventually arrives at the comb of
infinitely narrow δ functions, as in Fig. 6. On the other hand, if Nc is fixed (no
matter how large it is), with increasing s one eventually finds oneself in the energy
region where the resonance widths cannot be neglected; in fact, the resonances start
overlapping, and ρ(s) gets smoothed. The approximation of the infinitely narrow
resonances badly fails here, and must be amended. These two limits (Nc → ∞ with
s fixed, and Nc fixed with s → ∞) are not interchangeable. Later I will include

18
the nonvanishing widths (Sec. 10). For the time being I put Nc = ∞, so that all
resonance widths vanish.
The equidistant spectrum only holds if both, the primary and all daughter tra-
jectories, are exactly linear and parallel. This is not fully realistic. Even putting
Nc = ∞ does not help. The low-energy parts of the Regge trajectories in QCD
are not exactly linear, in particular, due to the spontaneous breaking of the chi-
ral symmetry and the emergence of the massless pions. One can explicitly check
that the radial excitations are not quite equidistant in the ’t Hooft model: they be-
come equidistant only asymptotically, at n ≫ 1 (n is the excitation number). The
situation in real QCD must be similar.
This shortcoming of the model affects the condensate expansion at low orders but
has no impact on deviations from duality at high energies. Since the phenomenon
under discussion is related to the factorial divergence of the high-order terms, letting
the low-lying excitations “breathe” does not change the factorial behavior [25], which
is in one-to-one correspondence with the spectral formula at large n, where the
2
spacings Mn+1 −Mn2 must be constant. Therefore, it is okay to use the linear pattern
(35). For the very same reason the residues of all resonances in Eq. (35) are taken to
be equal. Fluctuations of the residues would show up in the condensate expansion;
they do not affect the estimate of the duality violations. (For an additional remark
on the residues see Sec. 11.)
The infinite sum in Eq. (35) reduces to a well-known Euler’s ψ function, the
logarithmic derivative of Γ,

Nc Q2 + M02
Π(Q2 ) = − [ψ(z) + Const] , z≡ , (36)
12π 2 σ2
(see e.g. [26]). The irrelevant (subtraction) constant on the right-hand side is infi-
nite, strictly speaking. The occurrence of the Γ function reminds us of Veneziano’s
amplitude.
At positive values of Q2 an asymptotic representation exists for the ψ function,
∞
1 X B2k −2k
ψ(z) = ln z − − z , (37)
2z k=1 2k

where B2k stand for the Bernoulli numbers,

2(2k)!
B2k = (−1)k−1 ζ(2k) ; (38)
(2π)2k

here ζ is the Riemann function. (In some textbooks (−1)k+1 B2k is called the k-th
Bernoulli number and is denoted by Bk .) Equation (37) defines the asymptotic
expansion of the polarization operator,
∞
Nc Ck
Π(Q2 ) 2
X
−→
Q2 →∞
− ln Q + , (39)
12π 2 2 k
k=1 (Q )

19
where the coefficients Ck can be expressed through B2k (see Eq. (37)) in a relatively
straightforward manner. The leading logarithmic term exactly coincides with the
free quark loop of Fig. 1 presented in Eq. (5). This explains our choice of the
resonance residues. The next-to-leading term is 1/Q2 , followed by higher power
corrections. Equations (37) and (38) highlight the the factorial divergence of the
condensate series which I have already mentioned several times.
One might want to eliminate the 1/Q2 term from Π(Q2 ) to make the 1/Q2
expansion realistic – it is known [14] that, in QCD with massless quarks, the terms
of the first order in 1/Q2 do not appear in Π(Q2 ). The power series starts from
hG2 i/Q4 where hG2 i is the gluon condensate. One could eliminate 1/Q2 , say, by
fine-tuning the parameter M02 . If M02 = σ 2 /2 = 1/α′, the 1/Q2 term cancels.
Although this might seem desirable, in fact this is hardly worth the bother because
the model with exactly linear trajectories is too rigid to be realistic anyway. The
1/Q4 correction will come out way too large. To accommodate the gluon and mixed
condensates properly one would need a more flexible model, with more than one
adjustable parameter. Therefore, I will feel free to simplify the model further by
putting M02 = 0 and discarding the term with n = 0 in Eq. (35).
The spectral density corresponding to the infinite sum of the equidistant reso-
nances is shown in Fig. 6,
∞
s 2
 
σ2 = ′ .
X
ρ V (s) = ρA (s) = Nc · δ 2
−n ; (40)
n=1 σ α
Let us truncate the power expansion (39) at some finite order, and examine
the theoretical prediction for ρ(s) obtained from the Euclidean side. It does not
matter in which order we truncate. Any power term (1/Q2 )n is invisible in ρ(s)
at positive s: analytically continuing to positive q 2 and taking the imaginary part
one ends up with the δ function and its derivatives. The only imaginary part at
positive s comes from the analytic continuation of ln Q2 . Thus, in the model at
hand ρ(s)theor = 1. Comparing this with the comb of the δ functions in Fig. 6 we
conclude that the point-by-point duality is maximally violated: between the reso-
nances the spectral density is grossly overestimated (the “experimental” curve runs
below the “theoretical” expectation) while at the peaks it is grossly underestimated
(the “experimental” curve runs above the “theoretical” expectation). What is most
crucial, the deviations from duality do not die off with energy. The power index
vanishes!
All this is pretty trivial. A somewhat less trivial question worth examining is
the impact of the ad hoc smearing. For instance, let us average the spectral density
in Fig. 6 with the weight function appropriate to Rτ ,
I0 (Mτ2 ) I2 (Mτ2 ) I3 (Mτ2 )
Rτ = − 3 + 2 , (41)
Mτ2 Mτ6 Mτ8
where the moments In are defined as
Z M2
In (M) = ds sn [ρV (s) + ρA (s)] . (42)
0

20
 ρ (s)

15

12.5

10

7.5

2.5

2.5 5 7.5 10 12.5 15

s

Figure 6: The spectral density in the resonance model. For clarity I gave a tiny
width to the δ functions.

To estimate the oscillating contribution to Rτ which constitutes duality violation

that cannot be seen in a truncated OPE we treat Mτ as a free (large) parameter.
It will be seen momentarily that for Nc = ∞ and Mτ large, yet finite, the duality
violation in Rτ scales as 1/Mτ6 . In other words, the vanishing power index in ρ
translates3 in η = 6 in Rτ . Certainly, this distinction is quantitative rather than
qualitative, but, sure enough, it is quite important from the practical side.
The sum over resonances in Rτ is easily calculated analytically: for the spectral
density of Eq. (40) it is

Rτ = RτOPE + ∆Rτosc ,
!4
RτOPE σ2 1 σ2
= 1− 2 + ,
Nc Mτ 30 Mτ2
!3 !4
∆Rτosc σ2 1 σ2
 
2 2
= − x(1 − x)(1 − 2x) + x (1 − x) − , (43)
Nc Mτ2 30 Mτ2

where
Mτ2
!
x = fractional part of , x ∈ [0, 1) .
σ2

I presented the result as a sum of two functions of Mτ2 . The first one, RτOPE , is
a smooth function expandable in 1/Mτ2 . The second one, ∆Rτosc , oscillates with the
period σ 2 ; its average vanishes, see the plot of ∆Rτosc /Rτ0 in Fig. 7. Although ∆Rτosc
3
I remind that the power index η was defined in Sec. 3; in the case at hand ∆Rτ /Rτ ∼
Mτ−η sin Mτ2 /σ 2 .

21
 0.004

0.002

-0.002

3 4 5 6 7 8 9 10

Figure 7: Oscillations in Rτ . The plot of ∆Rτosc /Rτ0 is presented as a function of

Mτ2 /σ 2 .

is not a pure sine – it contains higher harmonics – the coefficients of the higher
harmonics are numerically suppressed. With the accuracy of a few percent one can
write !3
∆Rτosc σ2 Mτ2
!
1
=− √ sin 2π 2 . (44)
Rτ 3 12 Mτ2 σ
I pause here to make a comment. The contribution of any particular resonance
of mass Mk to Rτ , according to Eq. (32), is given by a simple polynomial in 1/Mτ2
(times the step function θ(Mτ2 −Mk2 ) ). Variations of parameters of a given resonance
(or resonances) change only the regular terms of the 1/Mτ2 expansion, but have no
impact on the oscillatory component. From Eq. (32) it is clear that such variations
change only coefficients of the 1/Mτ2 , 1/Mτ6 and 1/Mτ8 terms. OPE for Rτ must
exactly reproduce these three expansion coefficients, and so it does.
In fact, it is not difficult to demonstrate that RτOPE coincides with the OPE
prediction in the model at hand. The power corrections can be presented as follows:

I˜0 I˜2 I˜3

RτOPE = Nc + − 3 + 2 , (45)
Mτ2 Mτ6 Mτ8

where the “condensates” I˜n are

Z ∞
I˜n = ds sn [ρV (s) + ρA (s) − 2Nc ] . (46)
0

These integral representations for the “condensates” I˜n follow from Eqs. (32), (42)
if one assumes that the spectral densities approach their asymptotic limits faster
than any power of 1/s. In the model at hand, with the comb-like spectral density,

22
the integral representation (46) requires regularization. As a regularization one can
introduce the weight factor exp(−ǫs), taking the limit ǫ → 0 at the end. With this
regularization, RτOPE from Eq. (43) is reproduced.
The same result for the coefficients of the power terms in RτOPE could be obtained
directly from the expansion of Π(Q2 ) in Eq. (36). We agreed to put M02 = 0 for
simplicity. Then the expansion coefficients of Π(Q2 ) are those of the ψ function in
Eq. (37). To find the power terms in Rτ one may analytically continue the power
terms in Eq. (37) to Minkowski, take the imaginary part and convolute with the
weight function presented in Eq. (32). It is easy to see that the only relevant terms
in Eq. (37) are 1/z and 1/z 4 giving rise to δ(s) and δ ′′′ (s) in the imaginary part.
The terms of the higher order in 1/z drop out because the weight function is a
polynomial of the third order; it contains no s4 or higher terms. The term 1/z 2 in
Eq. (37) (it would generate δ ′ (s)) drops out because the weight function does not
contain terms linear in s. Then, the expansion for RτOPE is immediately recovered
provided that one substitutes the appropriate value for the appropriate Bernoulli
number, B2 = −1/30.
A few words on the numerical aspect. Our consideration is admittedly illustra-
tive. One should not take too literally the numbers which ensue for many reasons:
in particular, M 2 τ is not much larger than the spacing between the resonances,
Nc = 3 is probably not large enough to warrant the zero width approximation, and
so on. I would not put too much confidence on particular numbers. I would settle
on the statement that the above estimate of the oscillation component is valid, say,
up to a factor of two or so. Taking our formula for ∆Rτosc at its face value and using
the actual value of the τ mass (Mτ2 /σ 2 ∼ 1.5) we obtain that ∆Rτosc /Rτ ∼ 3%. It is
rather rewarding to see that this estimate is in the same ball park as that obtained
in the instanton-based model. It seems safe to conclude that duality violations in
the total hadronic τ width are expected at a level of a few percent. We do not
know whether two mechanisms add constructively or destructively. Given this ad-
ditional uncertainty, it will be no exaggeration to assert that the overall theoretical
uncertainty
∆Rτ /Rτ = 3% .
The 3% uncertainty in the hadronic τ width translates into ∼20% uncertainty in
αs (Mτ ), which entails the uncertainty of about 6% in αs (MZ ).

⋆⋆⋆

In summary:
• The resonance structure of the hadronic spectrum associated with the confining
properties of QCD leads to a distinct exponential/oscillating component invisible in
the truncated OPE.
• This component is related to singularities of the appropriate n-point functions
at infinite separations (in the coordinate space).

23
 • The corresponding duality violations are maximal at Nc = ∞, when the reso-
nance widths vanish. The power index which was introduced in Sec. 5 is calculable
in many instances. Generally speaking, the power indices obtained in the instanton-
based and resonance-based models do not coincide with each other. Rτ seems to be
an exceptional case. We do not understand the reasons explaining the coincidence
of the power indices in Rτ ; probably, it is accidental.
• Inclusion of the nonvanishing resonance widths will replace the power suppres-
sion in the pre-exponent by a weak exponential suppression, see Sec. 10. The change
of the regimes will probably have little numerical impact in the τ decays since Mτ2
is not large enough for the exponential regime to develop in earnest. The factor
(σ 2 /Mτ2 )3 in Eq. (44) will be replaced by
2πBMτ2
!
exp − .
Nc σ 2
Both factors are rather close numerically.

9 Numerical Illustrations in the ’t Hooft Model

The general pattern established in Sec. 8 is nicely illustrated by numerical cal-
culation in the ’t Hooft model which is ideally suited for exhibiting the resonance
mechanism of duality violations. Indeed, in the ’t Hooft model [23] (two-dimensional
quantum chromodynamics considered in the limit Nc → ∞, while αs Nc = const)
the color confinement is automatic because the Coulomb potential in 1 + 1 dimen-
sions grows linearly with distance. For high excitation numbers the meson spectrum
grows linearly, Mn2 ∝ n. In two dimensions the gluon field is a nondynamical degree
of freedom, there are no transverse gluons, and the only remnant of the gluon field
is the Coulomb (instantaneous) potential. There are no instantons, therefore, the
mechanism of Sec. 6 does not overshadow the resonance mechanism. Moreover,
the quark loops are suppressed at Nc → ∞ – hence, all quark mesons are infinitely
narrow. The spectrum is calculable, albeit numerically, and so are exclusive decay
amplitudes, which can be then summed up, one by one. This replaces the experi-
mental data, to be compared with the inclusive OPE-based calculations. In a sense,
this is even better than real data in actual QCD which are always incomplete and
imprecise. In the ’t Hooft model one can make gedanken experiments as complete
and precise as one wishes.
A thorough analysis of the quark-hadron duality in the semileptonic heavy flavor
decays in the ’t Hooft model was carried out by Lebed and Uraltsev [10]. Below I
will present their result, but at first let me remind relevant aspects of the ’t Hooft
model.
The dynamical mass scale in the model is set up by the gauge coupling constant
g (which is dimensionful in D = 2),
2 g2
β ≡ (Nc − 1/Nc ) . (47)
2π

24
Thus, β plays the role of ΛQCD . Being finite at Nc → ∞, it provides a natural unit
of mass. The quarks with masses less than β can be called light, while if mQ ≫ β we
are dealing with a heavy quark. To make contact with real QCD the heavy quark
will be referred to as the b quark, while the corresponding bq̄ bound state as the B
meson.
The semileptonic widths b → qeν̄ are induced by the weak decay Lagrangian
G
Lweak = − √ (q̄γµ b) (ēγ µ ν) , (48)
2
where G is a “Fermi constant,” and it is assumed that

me = mν = 0 . (49)

Note that in D = 2 the axial current is not independent, so that Lweak is chosen to
be pure vector times vector. Moreover, the invariant mass k 2 of the lepton pair is
always zero [8], provided the lepton masses vanish, Eq. (49). This fact is specific for
2D models; it implies that in two dimensions the semileptonic decays b → qeν̄ are
equivalent to decays into a single massless pseudoscalar particle φ weakly coupled
to quarks,
G
L̃weak = − √ q̄γµ b ǫµν ∂ν φ . (50)
2π
For simplicity it will be assumed that the quark q in the transition b → qeν̄
is distinct from the spectator light antiquark in the B meson, so that annihilation
diagrams are absent. The leading (parton-model) result for the semileptonic width
Γ(b → qeν̄) scales as m1b (to be compared with m5b in real QCD). A few first terms in
the operator product expansion for Γ are rather trivial; it is not difficult to obtain

G2 m2b − m2q mb 1 dx 2
Z
ΓOPE = · · ϕ (x) . (51)
4π mb MB 0 x B
up to terms terms O(m−4 b ). Here x is the b quark momentum fraction of B in light-
cone coordinates, while ϕB (x) is the light-cone wave function. The expectation
value (mb /MB ) h1/xi in Eq. (51) can, in turn, be computed in the form of a 1/mb
expansion.
Equation (51) presents a smooth function of mb , a theoretical part of the duality
relation. It carries no direct information on the opening of new thresholds each time
MB crosses the successive values of Mn where Mn is the mass of the n-th excited
state in the q̄q channel (I will denote the meson itself by the same symbol, Mn ).
The “experiential” part is obtained as a sum over individual exclusive decay
channels of the type B → Mn eν̄. Sure enough, the “experiential” part includes
the effects of the kinematic thresholds in the most straightforward manner. The
spectrum and the light-cone wave functions of the excited light mesons are found,
numerically, from the ’t Hooft equation [23]. The concrete calculations I would
like to quote are done [10] at mq = 0.56β. The b quark mass in these calculations

25
 0

-1

log10(ΓB/ΓOPE –1) -2

-3

-4

-5

-6

-7

-8
2 4 6 8 10 12
mb

Figure 8: Deviations from duality in the semileptonic B decays in the ’t Hooft

model, from Ref. [10]. The b quark mass is measured in the units of β, the light
quark masses are set at 0.56β.

varies from mb = β (unrealistically light, no phase space for excitations at all) up

to mb = 12β, when the highest kinematically allowed excited q̄q meson has the
excitation number 18. (12β corresponds, roughly speaking, to mb = 4.5 GeV in
actual QCD.) The result [10] for

(ΓB )“exp”
−1
ΓOPE
is shown in Fig. 9, as function of the b quark mass.
Seeds of oscillations inherent to duality violation are clearly seen. (The deviations
do not average to zero but rather oscillate around a rapidly dissipating contribution
which can be attributed to discarded higher-order OPE terms.) Numerically, the
duality violating component turns out to be extremely small – almost certainly, an
artifact of the ’t Hooft model. What I would like to emphasize is not the numerical
suppression per se, but a general character of the phenomenon.

10 The Impact of the Resonance Widths

As was already mentioned, no matter how large Nc is, at sufficiently high energies
one cannot neglect the resonance widths. One should also remember that in the
real world the number of colors is three – not very large by any count. One can
expect that the nonvanishing resonance widths, once the resonances start to overlap,
provide an additional suppression of the duality violating component.

26
 Below I will show that, inside the domain of overlapping, the power regime (10)
is replaced by an exponential one, see Eq. (11), even if the power index vanishes
(as is the case for the spectral density (4) presented in Fig. 6). The basic idea is
that the nonvanishing resonance widths shift the poles away from the physical cut,
to unphysical sheets, which automatically results in a smoother imaginary part on
the physical cut, much closer to that obtained from OPE.
I return to the equidistant resonances considered in Sec. 8 in the leading order
in 1/Nc, with the intention to include the next-to-leading 1/Nc effects. For any
given excitation n, the widths scale as Γn /Mn ∼ Nc−1 . Below we will analyze an
infinite sequence of resonances for which we will need to know the dependence of
this ratio on the excitation number n. For the time being, however, let us focus on
an “isolated” resonance.
In the leading order its contribution to the polarization operator Π(q 2 ) is
gn2
; (52)
q 2 − Mn2
here gn2 is the residue. If Γn 6= 0, in the vicinity of the pole one can use the Breit-
Wigner formula which replaces Eq. (52) by
gn2
. (53)
q 2 − Mn2 + iΓn Mn
Strictly speaking, gn2 and Mn2 in Eqs. (52) and (53) are different: in the former the
residue and the mass must be taken in the leading order in 1/Nc while in the latter
they include O(1/Nc) corrections. This effect is less important than that due to
Γn ’s. In many instances (but not always, see below) the 1/Nc shifts of gn2 and Mn2
can be neglected.
Now comes a crucial assertion. At large n the resonance width must scale as
Γn ∼ Mn /Nc , i.e. the n dependence of Γn is the same as that of Mn , the square root
of n. This behavior was predicted long ago [27] on the basis of a simple qualitative
picture; much later it was confirmed [6] by numerical studies in the ’t Hooft model.
This formula can be explained as follows. For highly excited states a quasiclassical
treatment applies. When a meson is created by a local source, it can be considered,
quasiclassically, as a pair of (almost free) ultrarelativistic quarks; each of them with
energy Mn /2. These quarks are produced at the origin, and then fly back-to-back,
creating behind them a flux tube of the chromoelectric field. The length of the
tube L ∼ Mn /Λ2 where Λ2 represents the string tension. The meson decay width is
determined, to order 1/Nc , by the probability of producing an extra quark-antiquark
pair. Since the pair creation can happen anywhere inside the flux tube, one expects
that4 Γn ∼ LΛ2 /Nc . Taking into account that L ∼ Mn /Λ2 one arrives at
B
Γn = Mn , (54)
Nc
4
Let me note in passing that the 1/Nc2 corrections
√ due to creation of two quark√ pairs are of
order L2 /Nc2 within this picture. Since L ∼ Mn ∼ n, the expansion parameter is n/Nc .

27
where B is a dimensionless coefficient of order one. Naively extrapolating Eq. (54)
to n = 1 and Nc = 3 (well beyond the limits of its applicability) and normalizing by
the ρ meson for which Γ/M ∼ 0.2, one can make an educated guess,

B ∼ 0.5 . (55)

In the ’t Hooft model, in which both the

√ meson masses and widths are calculable,
one can explicitly test the formula Γn ∼ n. It is curious that in the ’t Hooft mode
one arrives at a close numerical estimate for B. Figure 9 shows Γn versus n. The
width of the n-th excitation is computed numerically, by summing up all open two-
meson decay channels,
X
Γn = Γ(n → b + c) , Mb + Mc ≤ Mn .
b,c

The necessary three-meson constants are found [6] as appropriate overlap integrals
of the light-cone wave functions, the calculation of Γn extends
√ up to the excitation
√ curve yielding the n dependence fits well
number 500. It is seen that the solid
the numerical data.5 Since Mn ∼ n, the scaling behavior (54) is confirmed. The
coefficient B defined in Eq. (54) is obtained from the coefficient A of Ref. [6] as
follows: B = A/π 3 . Since A ∼ 15 (see the Erratum), we recover Eq. (55).
The Breit-Wigner resonance formula now takes the form (in the Euclidean no-
tation)
#−1
BMn2
"
2 2 2
− gn Q + Mn − i . (56)
Nc
One must exercise certain care in using the Breit-Wigner expression for the resonance
contribution in the polarization operator far away from the pole. It must be adjusted
in such a way that the analytic properties of Π(q 2 ) are not spoiled. Namely, Π(q 2 )
must remain analytic everywhere in the complex q 2 plane, with a cut along the
positive real semi-axis. No singularities are allowed on the physical sheet, all poles
must be shifted to unphysical sheets. This is relatively easy to achieve. To this end
let us replace Eq. (56) by
#−1
Q2
" !
B
− g̃n2 Q 2
1− ln 2 + M̃n2 . (57)
πNc σ

On the upper side of the physical cut near the pole (i.e. at Q2 in the vicinity of
−Mn2 − iǫ) the imaginary parts of both expressions (56) and (57) coincide, provided
that g̃n2 and M̃n2 are appropriately adjusted (g̃n2 and M̃n2 differ from gn2 and Mn2 by a
1/Nc correction; in what follows we will omit the tilde, making the adjustment at
5 √
Seemingly random fluctuations of Γn / n around a constant value are also clearly seen. A
theory of these fluctuations might be a good exercise. Developing such a theory might be fun.

28
 Γn

Figure 9: The width of the n-th excited meson in √

the ’t Hooft model, in the units
−2 −1
of 32π Nc β, from Ref. [6]. Note that Mn = πβ n at n ≫ 1. The parameter β
is defined in Eq. (47).

the very end). Thus, both expressions are equally legitimate for the Breit-Wigner
description of the resonances. Moreover, Eq. (57), in turn, can be replaced by
 !− B −1
Q2 πNc
− gn2 Q 2
+ Mn2  . (58)
σ2

The distinction between Eqs. (57) and (58) is of the order O(1/Nc2). We do not
track such terms anyway. This latter formula has the required analytic properties
– it is nonsingular everywhere in the complex Q2 plane, except the cut at negative
real Q2 . On the physical sheet of Q2 , the variable
!1− B
Q2 πNc
z≡ (59)
σ2

never becomes real and negative, so that the pole singularities in Eq. (58) are indeed
shifted to unphysical sheets.
Summing over the infinite chain of the equidistant resonances, as in Eq. (35),
we arrive at
 !− B −1
∞
Nc σ 2 1 Q2 πNc
Π(Q2 ) = Q2 + Mn2 
X
12π 2 1 − B/(πNc ) n=1 σ2

29
 Nc 1 1
 
= − 2
ψ(z) + , (60)
12π 1 − B/(πNc ) z
where z is defined in Eq. (59), and 1 − B/(πNc ) in the denominator reflects the
adjustment of the residues discussed above (this factor can be established by de-
manding the correct asymptotic behavior at Q2 → ∞, cf. Eq. (5)).
How does this compare with the zero-width approximation of Sec. 8? Formally
both results for the polarization operator look similar; the only difference6 is in the
definition of the variable z. In the deep Euclidean domain the power expansion
of Π(Q2 ) ensues from the asymptotic representation (37). The k-th term of the
expansion, which in the leading order in 1/Nc used to be
!2k
B2k σ2
(zero-width approx.),
2k Q2

now becomes
!2k
σ2 2kB Q2
!!
B2k B 1
1+ ln 2 + +O . (61)
2k Q2 πNc σ πNc Nc2

The impact of the next-to-leading 1/Nc terms on the power expansion is quite in-
significant. In particular, the (B/Nc ) ln Q2 correction mimics the logarithmic anoma-
lous dimension typical of OPE in QCD.
At the same time, the 1/Nc effects radically affect ImΠ on the physical cut (the
spectral density). Instead of the comb of the δ functions of Fig. 6, we now get at high
energies a smooth function, with mild oscillations. The plot of the spectral density
ρ(s) = (12π/Nc ) ImΠ (with Π determined by Eqs. (60) and (59)) is presented in
Fig. 10.
At asymptotically large s the spectral density tends to unity, plus power correc-
tions corresponding to OPE plus an oscillating component which cannot be obtained
in the power expansion. One can readily isolate it by exploiting the well-known re-
flection property of Euler’s function,
1
ψ(z) + = ψ(−z) − πcotπz . (62)
z
On the physical cut (in units of σ 2 )
!
B 1
 
B
1− πN
z = −s c 1+i +O , (63)
Nc Nc2

and the imaginary part of the left-hand side of Eq. (62) is essentially given by
that of −πcotπz. The imaginary part of ψ(−z) gives rise to non-oscillating power
6
The 1/z term which was absent in Sec. 8 is due to the fact that I put M02 = 0 and start the
summation from n = 1 rather than from n = 0. These simplifications resulting in the occurrence
of the 1/z term, are irrelevant in the studies of the duality violating component at high energies.

30
 2 ρ (s) B=0.5
Nc =3
1.5

0.5

s/σ2
0
1 2 3 4 5 6 7

Figure 10: The spectral density for the equidistant resonances with finite widths
specified in Eq. (54).

corrections s−2k in the spectral density corresponding to OPE. Alternatively, these

OPE corrections can be obtained by a direct analytic continuation to Minkowski of
the power terms (61).
As a result, the finite-width resonance model leads us to
2πsB 2πs
   
ρ(s) → 1 + power corr. + 2 exp − cos , (64)
σ 2 Nc σ2
where it is assumed that
2πsB ln s
≫1 but ≪ 1. (65)
σ 2 Nc Nc
If s is fixed while Nc is set to ∞ we recover unsuppressed oscillations, as in Sec.
6. In the opposite limit Nc fixed and s large we observe an exponential suppression,
with a weak exponent proportional to 1/Nc.
A key question one can ask in connection with the above analysis is as follows.
For the given value of Nc , what is the boundary energy marking the onset of the
exponential suppression? The best I can say at the moment is that this boundary
energy scales as s0 ∝ Nc . A reliable determination of the proportionality coefficient
is a task for the future. A naive estimate following from Eq. (65),

σ 2 Nc
s0 ∼ ∼ 2 GeV2 at Nc = 3 ,
2πB
if correct, would mean that the resonances essentially overlap (and, thus, smear the
spectral density), starting from the first or second excitation.

31
 5 ρ 11
00
00
11
00Aleph
11 data
Pert. theory
Instanton model
4
Resonance model

s, GeV 2
0 1 2 3 4 5 6 7

Figure 11: Experimental data (ALEPH) on the spectral density in the vector channel
and the duality violation models.

11 How Does All This Match Experiment?

In spite of the current extremely high demand on theoretical estimates of duality
violations, there is surprisingly little effort to elucidate the issue by direct exper-
imental studies. A high-precision measurement of the ratio R in e+ e− in a broad
range of energies, from threshold up to, say, s = 10 GeV2 in a dedicated experiment
with the proper absolute normalization would give an enormous boost to this issue.
Alas, such measurements have never been undertaken... The most accurate data on
the spectral densities were obtained in τ decays [28], where they (naturally) extend
only up to s = 3 GeV2 . This energy range is way too narrow to be helpful in per-
fecting the models which are currently in use or in the design of new models. Still,
confronting the data with current theoretic ideas might give a feeling of whether or
not we are moving in the right direction. The comparison is presented in Fig. 11. I
will first explain what is depicted, and then offer several comments.
The ALEPH experimental data [28] I show correspond to the spectral density
in the vector isovector channel. There exist some data above 3 GeV2 from e+ e−
annihilation, but the error bars are so large that plotting these points would just
obscure the picture. Directly measurable in the hadronic τ decays is the sum of
the vector and axial spectral densities. To obtain the spectral densities separately
one has to sort out all decays by assigning specific quantum numbers to each given
final hadronic state. In the majority of cases such an assignment is unambiguous.
For instance, two pions (whose contribution is the largest) can be produced only
by the vector current. Some processes, however, can occur in both channels, for

32
instance, the KKπ production. Using certain theoretical arguments it was decided
[28] that around 3/4 of all KKπ yield must be ascribed to the vector channel. Other
theoretical arguments [29], which seem more convincing, tell that virtually all KKπ
production takes place in the axial channel. Therefore, in fitting the data, I subtract
the KKπ yield from the data points presented in Fig. 11. I hasten to add that the
subtraction affects only the energy range s > 2 GeV2 , i.e. the second peak in Fig.
11, and even in this energy range the effect is small, <
∼ 5%. The subtraction is not
essential for a general picture I draw here.
The solid curve at s < 2.7 GeV2 is the best fit of the data points thus obtained
by the sum of two Breit-Wigner peaks (the first one, the ρ meson, is a modified
Breit-Wigner taking into account threshold effects important for the ρ meson). For
my purposes the two-resonance fit presents an excellent approximation to the ex-
perimental spectral density at s < 2.7 GeV2 .
The tails of the curves at 2.7 < s < 7 GeV2 represent the resonance-based and
instanton-based models (the solid and dashed curves, respectively) which I discussed
in Secs. 10 and 6. The solid curve is a modulation of ρ(s)pert by the factor

1 + 1.22 E −3 sin(2ρE − δ) ,

ρ = 3 GeV−1 , δ = 1.32 , E in GeV ,

cf. Eq. (18). The dashed curve is a modulation of ρ(s)pert by7

2πsB 2πs
   
1 − 1.24 exp − 2 sin − 3.08 ,
σ Nc σ2

σ 2 = 2 GeV2 , B = 0.5 , Nc = 3 ,

cf. Eq. (64).

Finally, the thick solid curve in Fig. 11 displays ρ(s)pert calculated through order
3 (3)
αs , with Λ = 200 MeV.
MS
It is clearly seen that, qualitatively, both models match the data well. In fact, in
actuality I would expect an oscillating component which is a combination of the solid
and dashed curves, since the mechanisms of the duality violations they represent are
complementary. The instanton mechanism leads to a visibly slower (power) fall off,
with rearer oscillations. It is also clear that experimental measurements at the
percent level of accuracy will most certainly provide us with the material needed for
construction of a reliable and well calibrated model of the duality violations.

7
As we know from Sec. 10, the equidistant resonances with equal residues result in oscillations
of this type superimposed on ρ = 1. By a trivial adjustment of the residues one can readily achieve
that the oscillations (64) are superimposed on ρ(s)pert .

33
12 “Exclusive” Duality
This section is somewhat perpendicular to the main theme of the review, and can
be safely omitted. While previously my topic was estimating duality violations,
now I will discuss a special “exclusive” mode of implementation of the quark-hadron
duality.
In all problems considered above duality is applicable at high energies (momen-
tum transfers), where the cross sections (decay probabilities) are saturated by a
large number of exclusive channels. Moving in the opposite direction, towards lower
energies (momentum transfers), we decrease the number of the open channels and,
typically, worsen the accuracy of the quark-hadron duality. There is a stereotype in
the public eye that “duality works well in inclusive processes after summing up over
a large number of exclusive channels.” Although this is often true, sometimes the
stereotype turns out to be wrong. In this section, concluding my review, I give two
examples when duality (i.e. “equality-with-controlled-accuracy” between the quark
and hadronic processes) takes place even though the hadronic process is saturated
by a single exclusive channel. Sure enough, such situations are rather exceptional
and are explained by certain custodian symmetries.
The first example of this type was discovered [30] in the mid-1980’s in the heavy
quark transitions at zero recoil. A prototype process has the form

Q → Q′ + ℓν (66)

or, at the hadron level,

HQ → HQ′ + ℓν . (67)
Here Q and Q′ are heavy quarks of distinct flavors, HQ stands for a Q-containing
heavy hadron (practically, the lowest-lying state, say HQ = B or HQ = Λb ), while
ℓν is the lepton pair. Needless to say that the sum over all possible HQ′ ’s in the
final state is implied since we consider the inclusive reaction. As we will see shortly,
at zero recoil, only one state will survive in this sum.
At the point of zero recoil, the total spatial momentum of the lepton pair van-
ishes, ~k = 0. Then the time component of the lepton pair momentum in the physical
decay (67) must be equal to k0 = MHQ − MHQ′ . The value of k0 is maximal if HQ′
is the ground state. It is convenient to introduce the notation
 
∆M = MHQ − MHQ′ , (68)
ground state

and
ǫ = ∆M − k0 . (69)
Then on the physical cut ǫ is real and positive.
Dynamical information is encoded in the differential probabilities which go under
the name of the hadronic structure functions wi (they were already mentioned in the

34
 k duality

Q Q’ Q HQ H Q’ HQ

(a) (b)

Figure 12: Inclusive semileptonic decay of one heavy flavor into another at zero
recoil. (a) Free quark transition operator; (b) hadronic saturation. The closed circle
denotes the insertion of the axial or vector current. The four-momentum of the
lepton pair is denoted by kµ .

very end of Sec. 7). One starts from the transition operator describing the forward
scattering of HQ to HQ via intermediate states HQ′ ,
Z
T̂µν = i d4 xe−ikx T {Q̄(x)γµ γ5 Q′ (x) , Q̄′ (0)γµ γ5 Q(0)} . (70)

For definiteness I limit myself here to the axial-vector current, Q̄γµ γ5 Q′ . (The vector
current can be treated in a similar fashion.) The hadronic amplitude is obtained by
averaging T̂µν over HQ ,
1
hµν = hHQ |T̂µν |HQ i . (71)
2MHQ
From kinematics we infer a general decomposition

hµν = −h1 gµν + h2 vµ vν − ih3 ǫµναβ vα kβ + h4 kµ kν + h5 (kµ vν + kν vµ ) , (72)

where vµ is the four-velocity of HQ (in the rest frame vµ = {1, 0, 0, 0}.) At zero
recoil hi ’s depend only on k0 . The physically observable quantities are the structure
functions
wi = 2 Im hi , 0 ≤ ǫ ≤ ∆M .
The relation between hi ’s and wi ’s is the same as that between Π(Q2 ) and ρ(s)
in e+ e− annihilation. Much in the same way as in the latter problem, one can
calculate hi far away from the physical cut, then perform analytic continuation onto
real positive ǫ, and then take the imaginary part, term by term in OPE. A new
element in the heavy flavor decay is that the expansion runs now in two parameters,
Λ/ǫ and Λ/mQ , where mQ is the heavy quark mass. It is assumed that Λ/mQ,Q′ ≪ 1.
Shortly I will send the quark masses mQ,Q′ to infinity keeping the difference mQ −mQ′
fixed.
At the quark level hi ’s are determined by the diagram displayed in Fig. 12(a).
By analogy with e+ e− annihilation one might think that, to make OPE meaningful,
it is necessary to require Λ/ǫ ≪ 1. Surprisingly, at zero recoil this hasty conclusion

35
is wrong. In fact, it is not difficult to prove (see e.g. [31]) that the structure of the
power expansion in the case at hand is as follows

Λ2
!
1
− h1 = 1 + 2 + ...
mQ − mQ′ − k0 mQ
Λ2
!
Λ
+ + ...
(mQ − mQ′ − k0 )2 m2Q

Λ2 Λ2
!
+ + ... + ... . (73)
(mQ − mQ′ − k0 )3 m2Q

If one takes the limit mQ → ∞ first, keeping finite the quark mass difference,

∆m ≡ mQ − mQ′ = fixed , (74)

then the expansion for h1 becomes absolutely trivial, namely

1
− h1 = (75)
mQ − mQ′ − k0
exactly, for all values of mQ − mQ′ − k0 , large or small. Thus, at zero recoil, the
free quark amplitude of Fig. 12(a) is exact, a manifestation of the heavy quark
symmetry.
Let us now examine h1 from the hadronic side (Fig. 12(b)). The very same
heavy quark symmetry tells us that the only state surviving in the sum over HQ′ at
~k = 0, mQ′ = ∞ is the ground state (Q′ q̄) where q stands for the light spectator
quark. Moreover,
F2
− (h1 )hadr = , (76)
ǫ
where F is the HQ → HQ′ form factor at zero recoil. Since ∆M = ∆m, both
expressions (75) and (76) perfectly match provided the transition form factor is
unity at zero recoil [30]. The quark-hadron duality is perfect: the quark pole of Fig.
12(a) is exactly equal to the hadronic pole presented in Fig. 12(b). Correspondingly,
the same equality takes place for the structure functions wi (k0 ).
The second example of “exclusive” duality I would like to mention is somewhat
more exotic. It was found [9] as a byproduct in the studies of the Pauli interference
in the heavy flavor decays in the ’t Hooft model. The setting has already been
presented in Sec. 9, which is devoted to the semileptonic decays. Now I pass to
nonleptonic decays. Correspondingly, the weak Lagrangian in Eq. (48) is replaced
by
G n 
¯ µ u + a2 dγ
 
¯ µ b (c̄γ µ u) + H.c.,
 o
Lweak = − √ a1 (c̄γµ b) dγ (77)
2
where I use the same notation as in Sec. 9; two distinct four-fermion operators in
Eq. (77), with (dimensionless) coefficients a1 and a2 , represent two possible patterns

36
 c

b b
d
B B
- -

u u

Figure 13: The quark graph determining the effect of the Pauli interference in the
nonleptonic inclusive B − decays at leading order in OPE.

of the color flow, direct and twisted (in actual QCD the latter emerges due to hard
gluon exchanges). At the quark level the Pauli interference in the B − decays is
described by the graph of Fig. 13,  which determines the coefficient in front of
the four-fermion operator b̄γµ γ u (ūγ µ γ 5 b) in OPE. To calculate the effect of the
5

Pauli interference in the quark language one has to calculate the diagram, take the
imaginary part and average the four-fermion operator over B − ,
1    1
hB −  b̄γµ γ 5 u ūγ µ γ 5 b  B − i = fB2 MB . (78)
 
2MB 2
If mb ≫ β, the result thus obtained has the same status as the parton-model formula,
say, for R(e+ e− ).
Following Ref. 9, I will set all quark masses except that of b to zero, mc = mu =
md = 0. Then the Pauli interference of Fig. 13 takes especially simple form
1
∆ΓPI (B − ) = − (a1 a2 )G2 fB2 MB . (79)
2
This effect is a 1/mb correction to the total inclusive B − width.8 At the same time,
∆ΓPI is of the same order in 1/Nc as Γtot (B − ). This is important.
As far as the hadronic saturation is concerned, the following statements are valid
to the leading order in 1/Nc :
(i) The intermediate states saturating ∆ΓPI are, by necessity, two-meson states.
Examples are shown in Fig. 14.
(ii) The graph of Fig. 14(a) is due to the first operator in Eq. (77) and is
proportional to a1 , while that of Fig. 14(b) is due to the second operator and is
proportional to a2 .
(iii) Factorization applies.
Generally speaking, both mesons in the intermediate state, π − and D 0 , need
not be ground states. Radial excitations would be acceptable, too. However, if
mc = mu = md = 0, the only hadronic states produced by the currents dγ¯ µ u and
8
In actual QCD the Pauli interference would be a 1/m3b correction; 1/mb versus 1/m3b reflects
the fact that the canonic dimension of the current q̄γµ q is one in D = 2 and three in D = 4.

37
 c
π-
d
b b
D 0

u u

B B
- -
c d
D 0
π-
u u u u

a b

Figure 14: The product of these two amplitudes determines ∆ΓPI to the leading
order in 1/Nc.

ūγ µ c are the ground state pseudoscalar mesons with zero mass. This is a special
feature of the two-dimensional theory, with no parallel in four dimensions. The fact
that (dū) meson in Fig. 14(a) is produced by the dγ ¯ µ u current forces it to be a
−
massless π , while, by the same token, the (uc̄) meson in Fig. 14(b) is a massless
D0.
As a result, the only intermediate hadronic state surviving in the Pauli interfer-
ence is
B − → massless π − + massless D 0 → B − (80)
This single exclusive channel must saturate (79).
So it does! This comes out rather trivially [9] because of the following properties
of the pion coupling and the B → D(π) transition form factors at k 2 = 0
s s
Nc 1 Nc
Z
fπ = φπ (x) dx = ,
π 0 π
s
1 1 π
Z
kµ εµν ¯ ν b|B − i = −kz
hπ − |dγ φB (x) dx = −kz fB . (81)
2MB 0 Nc

Here φ(x) is the light-cone wave function; for the pion φ(x) = 1. Note the occurrence
of fB in the transition form factor.
Using Eq. (81) it is easy to verify that the massless quark loop of Fig. 14(a)
perfectly matches the massless meson loop, the only contribution to ∆ΓPI surviving
at the hadronic level. The origin of the factor fB2 in Eq. (79) in the hadronic
calculation is totally different, though; it comes from the square of the transition
form factor, the matching seems magic, yet it could not have happened otherwise.
It is worth stressing that the magic simplifications which led to the “exclusive”
mode of the duality implementation in the problem at hand are explained by mirac-
ulous features of two dimensions.

38
13 Conclusions
In this review I identified general mechanisms causing deviations from duality and
derived scaling laws governing the damping of the duality violating component in a
variety of inclusive processes as a function of energy (momentum transfer). The esti-
mates of the absolute normalization one can perform at the moment are less certain.
This explains why I was so cautious in my discussion of the numerical situation in
Rτ , and completely avoided this issue in other inclusive processes. Getting a better
idea of the absolute normalization is absolutely necessary for all practical applica-
tions of the theoretical constructions presented here. Precision measurements of
ρ(s) in a wide energy range would be of enormous help in this question and would,
probably lead to a speedy solution.
As I have already noted, duality violations parametrize our ignorance. Were an
analytic solution of QCD found, the contents of this review would become instantly
obsolete. Then the exact asymptotic behavior of inclusive cross sections would be
known, and the very concept of duality violations would become irrelevant.

Acknowledgments
I would like to thank N. Uraltsev and A. Vainshtein for useful discussions.
This work was supported in part by DOE under the grant number DE-FG02-
94ER408.

References
[1] E.C. Poggio, H.R. Quinn and S. Weinberg, Phys. Rev. D 13, 1958 (1976).
[2] M.A. Shifman, in Continuous Advances in QCD, Ed. A. Smilga (World Scien-
tific, Singapore 1994), p. 249 [hep-ph/9405246].
[3] I. Bigi, B. Blok, M.A. Shifman and A. Vainshtein, Phys. Lett. B 323, 408
(1994) [hep-ph/9311339]. The ongoing debate on this issue, with sometimes
conflicting arguments, is reflected e.g. in M. Neubert, in Heavy Flavors II.
Ed. A.J. Buras and M. Lindner (World Scientific, Singapore, 1998) pp. 239-293
[hep-ph/9702375]; N. Uraltsev, in Proc. International School of Physics “Enrico
Fermi” Heavy Flavor Physics: A Probe of Nature’s Grand Design, Varenna,
Italy, July 1997, Ed. Ikaros Bigi and L. Moroni (IOS Press, 1998) pp. 329-409
[hep-ph/9804275].
[4] M. Shifman, in Particles, Strings and Cosmology, Eds. J. Bagger et al., (World
Scientific, Singapore 1996), p. 69 [hep-ph/9505289].
[5] B. Chibisov, R.D. Dikeman, M. Shifman and N. Uraltsev, Int. J. Mod. Phys.
A 12 2075 (1997) [hep-ph/9605465].

39
 [6] B. Blok, M. Shifman and D. Zhang, Phys. Rev. D 57 2691 (1998); (E) D 59
019901 (1999) [hep-ph/9709333].
[7] M. Shifman, Prog. Theor. Phys. Suppl. 131 1 (1998) [hep-ph/9802214].
[8] I. Bigi, M. Shifman, N. Uraltsev and A. Vainshtein, Phys. Rev. D 59 054011
(1999) [hep-ph/9805241].
[9] I. Bigi and N. Uraltsev, Phys. Lett. B457 163 (1999) [hep-ph/9903258]; Phys.
Rev. D60, 114034 (1999) [hep-ph/9902315].
[10] R. F. Lebed and N. G. Uraltsev, hep-ph/0006346.
[11] K.G. Wilson, Phys. Rev. 179 1499 (1969); Phys. Rev. D3, 1818 (1971);
K.G. Wilson and J. Kogut, Phys. Rept. 12, 75 (1974).
[12] This phenomenon is explained and illustrated in great detail in the collection
Large Order Behavior of Perturbation Theory, Eds. J.C. Le Guillou and J. Zinn-
Justin (North-Holland, Amsterdam 1990). One can find there all classical works
and an extended commentary.
[13] A. H. Mueller, in QCD: 20 Years Later, Eds. P.M. Zerwas and H.A. Kastrup
(World Scientific, Singapore 1993), page 162.
[14] M. A. Shifman, A. I. Vainshtein and V. I. Zakharov, Nucl. Phys. B147, 385
(1979).
[15] M.S. Dubovikov and A.V. Smilga, Yad. Fiz. 37, 984 (1983) [Sov. J. Nucl. Phys.
37, 585 (1983)].
[16] M. Maggiore and M. Shifman, in Proc. First Yale-Texas Workshop on Baryon
Number Violation at the Electroweak Scale, New Haven, CT, March 1992,
Baryon Number Violation at the Electroweak Scale, Ed. L. M. Krauss and Soo-
Jong Rey (World Scientific, Singapore, 1992), p. 153.
[17] For a recent review see T. Schäfer and E. V. Shuryak, Rev. Mod. Phys. 70, 323
(1998) [hep-ph/9610451].
[18] L. S. Brown, R. D. Carlitz, D. B. Creamer and C. Lee, Phys. Rev. D17, 1583
(1978) [reprinted in M. Shifman (Ed.) Instantons in Gauge Theories, (World
Scientific, Singapore, 1994), p. 168].
[19] R. D. Carlitz and C. Lee, Phys. Rev. D17, 3238 (1978);
N. Andrei and D. J. Gross, Phys. Rev. D18, 468 (1978);
L. Baulieu, J. Ellis, M. K. Gaillard and W. J. Zakrzewski, Phys. Lett. B77, 290
(1978);
T. Appelquist and R. Shankar, Phys. Rev. D18, 2952 (1978);
M. S. Dubovikov and A. V. Smilga, Nucl. Phys. B185, 109 (1981).

40
[20] M. Shifman, “Lectures on heavy quarks in quantum chromodynamics,” in ITEP
Lectures on Particle Physics and Field Theory, (World Scientific, Singapore,
1999), Vol. 1, p. 1 [hep-ph/9510377].

[21] V. Novikov, M. Shifman, V. Zakharov, and A. Vainshtein, ABC of Instantons,

in M. Shifman, ITEP Lectures on Particle Physics and Field Theory, (World
Scientific, Singapore, 1999), Vol. 1, p. 201.

[22] B. Blok, L. Koyrakh, M. Shifman and A. I. Vainshtein, Phys. Rev. D49, 3356
(1994) [hep-ph/9307247].

[23] G. ’t Hooft, Nucl. Phys. B75, 461 (1974) [Reprinted in G. ’t Hooft, Under
the Spell of the Gauge Principle (World Scientific, Singapore 1994), page 443];
see also F. Lenz, M. Thies, S. Levit and K. Yazaki, Ann. Phys. (N.Y.) 208,
1 (1991); C. Callan, N. Coote, and D. Gross, Phys. Rev. D13, 1649 (1976);
M. Einhorn, Phys. Rev. D14, 3451 (1976); M. Einhorn, S. Nussinov, and E.
Rabinovici, Phys. Rev. D15, 2282 (1977); I. Bars and M. Green, Phys. Rev.
D17, 537 (1978).

[24] A.B. Kaidalov, Usp. Fiz. Nauk 105, 97 (1971) [Sov. Phys. Uspekhi 14, 600
(1972)];
P.D.B. Collins, An Introduction to Regge Theory and High Energy Physics
(Cambridge Univ. Press, 1977).

[25] A. R. Zhitnitsky, Phys. Rev. D53, 5821 (1996) [hep-ph/9510366].

[26] Z.X. Wang and D.R. Guo, Special Functions (World Scientific, Singapore, 1989).

[27] A. Casher, H. Neuberger and S. Nussinov, Phys. Rev. D20, 179 (1979).

[28] R. Barate et al. (ALEPH Collab.) Z. Phys. C76, 15 (1997).

[29] S. Eidelman, L. Kurdadze, M. Shifman, and A. Vainshtein, 1997, unpublished.

[30] M. A. Shifman and M. B. Voloshin, Yad. Fiz. 47, 801 (1988) [Sov. J. Nucl.
Phys. 47, 511 (1988)].

[31] I. Bigi, M. Shifman, N. G. Uraltsev and A. Vainshtein, Phys. Rev. D52, 196
(1995) [hep-ph/9405410].

41